Transcript for "The CX leader's playbook for voice AI: Practical strategies you can steal to transform phone support": Hello. Hello. Welcome. I'm gonna wait a couple minutes to just make sure we get a couple folks in. But feel free to chime in in the q and a. Maybe let us know where we're you're coming from. If you've just joined, welcome. Hello, James. Yay. Getting the party going. Thank you. Appreciate it. Great. Okay. I'll give everyone a minute to to just, kinda enter here. Welcome. Welcome. One more minute. Where? Oh, hello. Hello. Roswell, Georgia. Fantastic. Vancouver, Washington. Love it. Texas. Great. We, on the assembled side, are joining from foggy San Francisco, so very jealous of anyone in a in a great climate right now. Seattle, Washington. Awesome. Great. Hello, everyone. Alright. Why don't we get started? So good morning or good afternoon, and welcome all to our live digital event, the CX leaders playbook for voice AI. Thank you so much for joining us. We really appreciate it. My name is Cassandra Cassandra. I'm a product marketer at Assembled, also serving as your emcee today. We have a packed agenda, so I'll just jump into some introductions here so we can get started. Today, we're joined by May Cromwell and Sam Carlson, our voice AI experts from Assembled. Assembled is the all in one AI platform for superhuman support that combines intelligent workforce management and omnichannel AI agents so support teams can operate efficiently and at scale without ever sacrificing quality. We're also thrilled to have Leslie Ong here joining us from FlexCar in Boston, Massachusetts. FlexCar is a flexible alternative to car ownership, offering the world's first ever month to month car lease. We'll kick off today's session with May and Sam, who will chat through the why and how of voice AI today along with some real world examples. Then May and Leslie will continue that conversation in a fireside chat where Leslie will share a bit more about her experience getting VoiceAI up and running at FlexCar. Finally, we will end the session today with a bit of audience q and a from you all. So if you do have questions throughout the session, please go ahead, pop them in the q and a on the screen to the right, rail on your screen right there. We'll make sure to get to them at the end. So feel free to fire them off throughout the session. And without further ado, I will hand it off to Sam. Alright. Hello, everybody. Thanks, Cassandra, for the great intro. And first of all, just super excited to be presenting to everyone here and really looking forward to getting into our content today. We wanted to start it off with just a quick overview of the current state of phone support. So despite the addition of all kinds of new support channels over the last couple of decades, good old fashioned phone support remains the most critical channel for many businesses. It's the preferred way of contacting support for most consumers. 44% of consumers prefer phone support, trailed by 1715% for chat and email respectively. Despite being the most online generation, Gen z continues to use use phone support as much as any other channel. And support leaders report that call volumes continue to increase year over year by as much as 20% or more. And finally, current technologies, especially sort of phone trees or IVRs, provide generally a pretty poor customer experience and oftentimes pretty limited resolution rates. The good news is in the past couple of years, generative AI has really transformed what's possible with voice technology. So multiple technological breakthroughs have happened in just the last couple of years. Text to speech models can now produce human like, speech in real time. AI reasoning models can carry on a dynamic, intelligent, and helpful conversation with a customer, And improvements in speech recognition and turn detection enable these systems to handle interruptions and background noise better than ever before. So, really, for the first time, we can build truly conversational AI agents, which can provide dynamic and intelligent interactions with, low enough latency to provide a genuinely, enjoyable automated voice support experience. And even the most advanced, like, rules based systems from the early twenty twenties, will start to feel antiquated pretty soon. So great. You know phone support is important. You know AI is changing the customer support landscape. So what should we actually go do about it? And first, we wanted to talk a little bit about your AI strategy. If you break down your phone support operation, there are a bunch of things that AI can handle well mixed in with about as many things that AI can't handle well at all. So to successfully adopt AI, first, you have to understand the difference. In terms of things that AI does well, first off, we have issue intake, classification, and routing. So AI can be your first point of contact and conversationally clarify and classify the, caller's issue. It can collect the right intake information and route the case to the right place, whether that's the right agent, an automated resolution, or just creating a follow-up for your team if no one's available. The AI is infinitely scalable and can handle all of these touch points very accurately and efficiently. The second bucket here we have is knowledge based resolutions. We found that 20 to 30% of phone support cases can usually be solved using knowledge only, meaning that the case can be resolved without taking any actions. There are cases these are cases that look like, answering a question about a refund policy, explaining how to access something in your application or asking about which specific amenities are available at which specific location. AI can easily process all of your written down knowledge and answer these questions instantaneously and conversationally. And the third bucket here we have is well defined workflows. So another significant portion of case volume looks like common cases that can be solved with well defined repeatable workflows using access to back end systems. These are cases that look like processing a refund, resetting a password, or looking up the status of a payment. With the right guardrails and system access, AI can also effectively and securely resolve these cases as well. So this brings us to the bucket of things that AI doesn't handle quite so well. First of all, is any case that really requires a personal touch. Many support issues just simply can't be resolved by a bot in the same way that they can be by a human. So for any sensitive cases, situations where customers are upset and the situation needs to be deescalated or when building, like, a customer relationship or trust is really critical, member of your support team should handle the case. The other bucket here is just complex problems when things go kind of off the rails. So when your systems might have incorrect information or your knowledge base doesn't have any relevant information about the issue that someone is experiencing, a human is typically needed as well. AI systems are usually only as good as the data that they have available to them, and most companies have situations, at least a few, where the data isn't exactly perfect. So in these cases, AI just needs to recognize that a human touch is required and quickly escalate the call to the right agent and importantly, ensure that agent has all the context they need to pick up where the AI left off. When you get this right, it's kind of a situation where everyone wins. Customers with simple issues get high quality phone support instantaneously with no wait times. Customers who need more complex or sensitive support can quickly route into the right agent without having to navigate a frustrating phone tree or wait on hold for as long. Agents get to spend more time helping customers when it's most needed and less time on repetitive tasks, and companies provide great support for less cost. Now that we've talked a little bit about the overall AI strategy, we want to go a little bit deeper into what this looks like in practice. So first, starting off with sort of AI driven resolutions. As mentioned, there's kind of two ways that your AI agent can resolve cases over the phone. The first is knowledge based answers. By using your existing knowledge base to automatically respond to a wide range of customer issues, You can really quickly get to that sort of 20 to 30%, automation rate. And a platform like Assembled can automatically integrate with your knowledge base across a bunch of different options so you can have these up and running in a matter of days. The second bucket is these multistep workflows. So by giving AI the instructions and system access it needs to properly resolve cases that require system actions, you can increase the resolution rate to that set of cases that require actions. A platform like Assembled gives you a local builder to easily define those steps and policies for correctly handling these cases as well as a robust integration platform to connect all your back end systems. And for everything you can automate, again, AI can help you with just that intake classification and routing piece. So as your first point of contact, AI can conversationally classify the caller's issue and sentiment. It can then decide whether to attempt to resolve or hand off the case. It will capture intake information automatically, just saving your agents time when they do pick up cases. And it will hand off to your team with full context so the agents can pick up easily where the AI left off. With Assembled, all these settings are highly configurable, so you get to decide what types of cases to resolve and which ones to hand off straight to an agent, whether to transfer to a live agent or just create a follow-up ticket for the issue, or what types of info to capture via, AI before handing off to your team. So with that, I'm gonna pass things over to May Cromwell who will talk a little bit more about the process of implementing AI for phone support and walk us through some real customer examples. Cool. Thanks, Sam. Yeah. So I wanna chat a little bit about implementation. And a typical implementation here at Assembled is pretty quick, usually around six weeks. And we've found the most successful rollouts are really supported by a dedicated AI deployment strategist along with a product team that's really focused on bringing you the latest and greatest in voice AI. So we usually kick things off by aligning on a customer's goals. So, really, what are they trying to automate, while also integrating with their key platforms? From there, we'll spend a few weeks on the setup. So that means configuring your AI agent and really building out your most common and high priority use cases. And then finally, we really test and iterate based off of initial feedback, before ultimately going live, and really gradually expanding to a broader set of use cases. So the value just continues growing over time. Now when it comes to tracking AI agent success, there are a few key metrics we recommend keeping an eye on, which I'll talk about on the next slide. And these are so at the top of the list is resolution rate, abandon rate, and handoff rate. And then across both AI and human agents, it's really important to monitor handle time, CSAT, and QA scores. And here you'll see a few images. These are just a few example dashboards within Assembled that really help teams understand not just how their AI agents are performing, but also how their human agents are performing as well and really where there's room to improve. Now let's look at how this plays out in the real world with a few, real life customer examples. And I'll keep these examples relatively brief just so we can get to our live conversation with Leslie from FlexCar. But first up, we have a fintech company, that was really struggling with its legacy IVR. Customers were stuck in the IVR for really over a minute, often routed to their most expensive agents who oftentimes could not actually resolve the issue and just had to provide reassurance to customers. For example, like, a bank transfer might take a few days. And this is really where voice AI shines. So it can collect information naturally, through conversation and really route calls more accurately and efficiently. So when those conversations are handed off to a human agent, they're also armed with more context. What that customer said, what are the tailored next steps in that specific example. So really no one has to repeat themselves and the experience feels seamless. And then on the next slide, I'll talk about another example of an ecommerce company, with a familiar pain point. They had a goal of automating 40% of their phone volume, but their legacy voice agent just was really not quite hitting their ceiling. They couldn't handle the complexity of their returns and exchanges, which made up about 20% of their volume. So they turned to Assembled Voice to really tackle those more complex returns, among a bunch of other use cases while also just giving customers that smoother, more conversational experience. And with that, I'm excited to turn it over to our customer spotlight, which is Leslie from Flex Car. Leslie, would you mind just telling us a little bit about Flex Car and your role there? Oh, you're on mute still. Sorry. Yeah. Sure. So Flex Car is a flexible month to month car subscription service. Think of us like a car lease but without long term commitment, big or big down payment or being stuck with just one car. Our members can actually choose to swap into different cars whenever they feel like it. So right now, we're currently in New England, Atlanta, Charlotte, and Nashville, and we're actively looking for our next city. When I joined, Westcar three years ago, my focus was really just building out the WFM team to ensure that our customers could get timely support. Over time, I, took on, building out tools for our agents and also just leading the voice of the customer program, Basically, making sure that the feedback that we're getting from customers flows through to our product marketing and finance team so that we can continue to improve. Customer experience is like everything at Flexcard because we know that we're a new product, so people naturally have questions about how it all works. We're also still building our name. We're just finishing our third year officially. So we're still needing to build trust with our consumers. And then, also, we know that getting a car is not a commodity I commodity item, and, like, it is something that people think a long time about. So being able to get, transportation that is reliable and to get where you need, it is essentially part of someone's, like, livelihood. Right? So while we are one of the more affordable options to having a car, we know that it's just it's not a small ticket purchase and, like, customers are constantly evaluating us even when they're are. Totally. Yeah. Cars are such a critical part of a lot of folks' lives and really important to be able to provide a lot of that feedback back to your product team. I'd love to hear a little bit about kind of your current architecture of your care team, as you call it at FlexCar. So what does your support stack look like, channels, technologies, those sorts of things? Yep. So for our systems, we're using Assembled for WFM, Guru as our knowledge base, and then Maestro for QA. For our phone support, we're using 59 for IVR and also just getting the call to the customer. And then those will integrate into Zendesk and create tickets, which we also use for email and also, we create tickets for processing orders too. Most customer data like billing plans, customer rewards, like any of these nuances to their profile, those actually live outside of Zendesk, in our homegrown customer hub. And then in terms of staffing, it is interesting because our customer base right now is, 7,000 active customers month over month, and our support team is really quite small in comparison. We have 18 to 20 agents during the day between eight to five, and they handle all of the contact, both phone and email, and at some point, chat in the past too. We have four agents that we rely on for overnight support because as a car company, we need to, provide our customers with twenty four seven roadside and emergency assistance. That's super helpful. I'm curious to hear maybe a little bit about what limitations in your existing phone support and system really made voice AI a priority for you all? So, yeah, in the next year, we are expecting a ton of growth, but, you know, like, we just don't have the time to to plan and to stop for it. So we're looking to expand to cities where maybe we aren't in a household name yet, and we're expecting to see a spike in questions about just how everything works. So the challenge is that we wanna grow, but and we wanna engage more customers, but we don't necessarily want to grow our headcount at the same pace. And this is, like, the typical growing hurdles, like, just the cost of adding new headcount to the org, the time for training and onboarding to make sure that human agents know how the product works and that they found convincing on the phone. Also staffing for occupancy. We know that, like, if you have a volatile forecast like us, you occupancy is either going to be really expensive or you're risking burnout for agents. The last factor is just efficiency when we're getting new hires up, up to speed on, like, handle time and all those traditional call handling metrics, we still bake in follow-up times where we rely on a human to take a note and to perform these manual steps or manually escalate a a call over that ends up, you know, just being inefficient. And, also, right now, like, 60% of our, channel mix is phone. So all the work that we've been doing with Assembled for, like, AI on email, that isn't yet translating to, the bulk of our volume. And then moving back to the growth problem, we know that 30% of our contacts right now are from new customer that are just engaging with us, and we're having to prioritize, like, what calls our human staff is handling. And then lastly, with our business hours being really short, eight to five, we're aligning that because that's typically when, like, all the shops are open and all of these other business partners are open, but we know that customer shopping trends are different. Typically, people will start thinking about these decisions after work or when they have free time or even late at night. So right now, like, we do feel like we're missing a lot of opportunities to engage with those customers. Right now, in our traditional IVR, the the only option that we really have is to tell the customer to try again later or to write us an email. We feel like some of these customers drop off because of that. And then the last thing is that, like, we're just looking to give customers that, current customers that need support at night another outlet for us to to, be able to help them. Mhmm. Totally. And it sound I think you touched on this a little bit, but I'm curious how did you really kind of identify and prioritize some of those use cases you talked about, whether it's new customers, or folks that are reaching out outside of those normal hours? Yeah. So what we found with the traditional IVR is that we've always struggled to get, like, proper contact driver reporting. And to some of the points that Sam mentioned earlier, there's always a margin of error when we're relying on either the customer or agent to pick what the the reason is from a list that is preset. So to fix that, like, that's why we're also exploring using AI in front of our IVR to just simply ask the the customer in the greeting, what can I help you with? Thanks for calling Westcar. Is there do you have any questions that, we can help you with today? And then, we'll from there, we'll either have the AI workflow, try to get the customer to, their resolution faster, or, like, we will hand off to the right skill. And then right now, I feel like also with, with humans taking notes, like, we lose accuracy on what the call was for, and we're hoping that with AI, we have more accurate interaction summaries. Mhmm. Totally. That makes a ton of sense. And I know we talked a little bit about kind of best practices for developing that AI strategy. Curious how FlexCar really developed that strategy and what your approach has been. Yeah. So our ultimate goal is to make sure that we're using AI in a way that enhances the customer experience and gets the customer, solution faster versus acting as a sort of barrier to help. I think in the, like, the early stages when we're looking at, like, text to speech or, some sort of, like, bot system, it it just create a frustration for customers. And we know that, like, the AI agent is only gonna be as good as, like, what we allow it to do and what we train it to to do. So we have engagement with our product and our tech team to make sure that, like, we are creating that data orchestration layer that needs to happen. And then we also took phase a phased approach to this. So last year, we really focused as, AI as, like, an agent assistance tool to make our agents faster. And it was a lot lower stakes. Like, there was always supervision. And then now this year, we're starting to use AI for, or, actually, like, we have already started using AI to work silently, like, alongside the agent to escalate these contacts to the teams that need to see, like, what the customer experience is. And then they're able to jump in and do, like, the very specific things or start making systematic improvements to our, like, policies and workflows. This year, we do feel like we are ready to use AI to handle contact and try to keep these contacts contained so that that human intervention isn't necessary anymore because we saw that for the most part, they are just telephoning the message back to the customer, and there's no other action required except giving the customer the information. Mhmm. That makes a ton of sense. I'm curious, like, how did you really make a case for that change, especially with some of the leadership and executives at Flex Car? Yeah. So I was lucky because we started pushing these three tenants for customer service even before AI, when we basically needed to get things done for the care team and get the support of, like, product and leadership. And what we found is that those three tenants, like, still hold true and even made a stronger case when AI became an option. So the first thing was that we wanna make sure that our agents are handling the right contact. So these are contact where the human touch is necessary and it adds value. And we knew that we also had to push self-service as much as we can. Before AI, this was, like, the old school chatbot or, you know, pushing them to basically read the help center articles or even adding, like, prompts in the IVR where the outcome was just the customer thing, speak to a representative over and over again. So we knew that, like, that wasn't great, and we took that offline pretty quickly. And then the last thing was we wanted to take timing out of it. We wanted to change how urgently we had to answer these contacts so that, like, we didn't have to deal with typical WFM metrics like, like, ASA, like, occupancy, like, making sure that all calls were answered in thirty seconds and having the right staff to do that. So before AI, we did that with queue callback campaigns and, like, pushing those contacts over to email. But even when we did that, it still required a human agent to eventually handle those calls. And then, the last thing is that, like, when we started using some versions of AI, like, it became super apparent where where we can adopt voice AI. Because in these workflows, the the agents are they're looking at the knowledge base, and then they're saying the same script over and over again. Mhmm. Totally. I think we've seen definitely AI spike in some of that individualization and empathy pieces, which is really helpful when a lot of agents are kind of used to those same scripts and narratives. Could you walk us through a little bit of your evaluation process for AI and kind of what were the capabilities or features that you were really looking for? Yeah. So, we were looking for AI that we were confident that we could keep a high containment rate with. So, like, we'll be looking at how many times the customer still needs to be transferred or or if they're requesting to be escalated. And also the time to resolution for those contacts because we know that some of these things that these customers are asking for, they require, like, more complicated workflows. So that ticket, when it's going through our human teams, might take two or three times that we're hoping to bring down to. We're also looking at correct case tagging and, like, resolution path to be able to really define all the different scenarios. We're looking at, like, accurate sentiment analysis to gauge, like, what customers are frustrated about, what makes them, like, super happy. And, like, we think that with the sentiment analysis, we can do a lot of really personalized and exciting stuff for customers. And then the last thing is and the reason why we're, you know, exploring AI in general across the board for all parts of the company is that we're trying to get more done. So for care, we're looking at new volume that we're able to now take by expanding our hours without necessarily staffing all of those midnight hours. Mhmm. Totally. And I'm curious how to kind of some of those capabilities and features translate into how you're really measuring the success of your voice AI implementation and specific kind of KPIs you'll be tracking? Yeah. So, you know, ultimately, our goal is to just reduce the need to continuously grow our staff. Be but we want more customers. Right? So, we'll start by building, like, a percent of contacts that Cal can help us handle, into our WFM requirements. So we're saying that we know that we can absorb this many new customers, but this much probably won't require a human intervention. And then also our most like, our our first test for ROI that we're trying to get to is destaffing the overnight team and allowing customers to self-service, through a combination of voice AI and also, like, their engagement with our app. The overnight schedule is actually really tough for human agents. They're waiting for such a small number of calls, so they could be waiting all night for maybe four calls spread over eight hour shift, and it is really it's it's hard for our human agents too. Mhmm. Yeah. I'm curious if you're kinda making any changes in your support team structure. I know you mentioned kind of the overnight team, but any other kind of accommodations with that introduction of voice AI? Yep. So as we start using voice AI, we'll still be handing off these high touch critical cases that in that matrix, it makes more sense for a human to handle right now. But we are looking for to to automate some things that are more standard. For example, like submitting an incident claim or just looking up status of something on someone's account, like a status on a refund. For aspirationally, we're also looking at using voice AI in combination with our app for very specific things like setting up a vehicle appointment, or triaging, like, simple questions that, like, an SME typically needs to know, like car questions. I don't know anything about cars, but I'm sure that we could train the AI lot to answer perfectly. Okay. Awesome. Can't say I know a lot about car mechanics either, so that makes a ton of sense. And then curious, like, you mentioned, you know, before that occupancy is one of the most important call center metrics. Can you tell us a bit more about that specifically? Yeah. So I think with occupancy, we're it it's just hard to predict, especially if you have a you have a, very volatile forecast. Right? So, like, we know that occupancy during the day, during our poor hours, is super high. But at night, like, we're basically overstaffing with super low occupancy for the for a small amount of critical, critical call. And then in order to basically make sure that our our customers, that our agents aren't occupied when those critical calls come through. We're, like, changing the way that customers can contact us and only funneling those calls through, which even lowers occupancy at that point further. But, like, with AI, occupancy is not really a metric that we even have to worry about. Mhmm. Totally. Cool. And then maybe just looking ahead a bit into the future, curious how you see, you know, voice AI evolving at FlexCar in the next year or maybe even two years. Are there kind of, like, additional use cases you're thinking about beyond that initial automation? Yeah. We're thinking about using voice AI from different points in our app to to actually, like, bring it in so that it's part of, like, our service. So, very specific use cases that I just mentioned. But, you know, like, we can trigger this. Like, the opportunities are really endless. Like, it could be triggered from any point in the app. It could be, even, like, let's say, like, they want to initiate something through through chat or through the app. It We're not looking at voice AI as just on the phone, but, actually, like, how do how do the how do our customers engage with our app and our knowledge base in general. Mhmm. Totally. And then what kind of advice would you give to CX leaders who are just beginning to really think about how to mod modernize their phone support with AI? I think it's worth an investment. It's definitely gonna take some time to to get everything set up, and that's really that that was more on our part than Assembled part where we had to build out this knowledge base. But once you do, you really quickly see the the ROI on it. And, like, both from staffing and from the customer experience. And then, like, I think we all agree that traditional IVR models and, like, that early speech to text model is frustrating for the customer. They're just smashing numbers on their phone. And then what we also found was that for a startup like us where things are constantly changing, making sure agents are consistent, when they're looking up guru articles and also just understanding that knowledge, in real time, like, fast enough so that sounds natural and sounds good on the phone is is much harder than email. So it makes sense to apply AI more so for that as an you know, either just taking over the call or as an agent of this tool, even more so than just for emails. And, also, it helps keep that brand voice, cohesive across all channels, which is something that we were struggling with. Mhmm. But I think the reason that I feel the most strongly about moving to modernizing just everything in in care with AI is that, like, I feel like most companies, they will have teams that do all of this quantitative analysis to see how the customer is doing, and, like, what their, you know, churn likelihood might be and all of that. But what we know is that, like, what drives the customer to really rave about a product or or maybe even to leave a bad review is is less predictable. It's more emotional. And, like, before AI, it required very manual processes to get those insights. But now with AI, we can leverage, like, all of the data that we're getting from those customer interactions. And then, like, us as a customer experience leader, we then have the keys to all these, like, valuable insights that we can then share with the business. So I think it fundamentally will change the role that, the role and the impact that the customer experience has in, like, all business operations. Mhmm. Totally. That makes a ton of sense. Maybe a last question I have is, you know, looking beyond just phone support, curious what AI capabilities are on your support road map either this year or in following years. So I think that one of the things that we're keen on seeing how it plays out is how well AI can handle real time, like, image analysis. So, like, for us, like, our use case would be something like if a customer reports that they were just in an accident and they needed to submit their picture. They'll they'll you typically fill out a form or fill fill out a form or they'll text us in a picture of the damages. We'd like to be able to kick off those workflows that we have, like, with AI to kind of, like, give the customer an expectation of what the next steps are. But I think image analysis is in image analysis and also omnichannel AI, like, when we talk about, okay, like, the start of the phone call, how do we pick up that conversation after the handoff? I think that that's what we're looking for. Awesome. Cool. Well, thank you so much, Leslie, for taking the time, and chatting everything about Flex Car and all the interesting things you all are doing over there. We at Assembled definitely love working with your team a lot. I will pass it back to Cassandra to start talking through some of the questions. I know there have been a ton of questions in the chat, and then we can chat through those together. Yes. Great. Thank you so much, Leslie. I really enjoyed what you mentioned about, you know, thinking about voice AI in total beyond the phone channel as, like, a driver of insights. Really interesting. Appreciate your your insight there. So really excited to open up the floor to questions for the team. We've got a ton. Thank you so much for your engagement in the chat. I'm gonna go through them, and, we will answer in order. So our first question we have from Mary Rose is, how do you authenticate the caller before providing responses to ensure the security account? This might be a great question for Sam. Yeah. Happy to answer this one. So the short answer is that this kinda works like other voice, automation systems you've seen. So we will connect to your authentication system usually just via API. And, usually, the way this works is we will automatically hit an API based on the phone number of the person calling, pull out sort of some key account information, and then request that information from the user over the phone and verify that it's a match. So it's like a pretty standard framework for voice authentication, and something that we support and Assembled AI voice agent as well. Killer. So next question here. Have you found any challenges in integrating APIs into voice AI channels? If so, how did you overcome them? And maybe this is a good, question for maybe Sam and Leslie, I'd say, either of you. Do we have we found challenges there on our side with with all our work, assembled with customers here? I'm happy to take this one, but, Leslie, jump in if you if you have anything to add. So I think with connecting to APIs, there's a few considerations that are just specific to voice. One is the latency of the API. So, usually, with a voice agent, you're trying to keep response time very quick. And some API requests can take longer to complete and then, like, parse the response and use that to respond. So one thing we do is sort of queue to the caller that it'll sort of be a second. And in the same way that a human would when looking up a piece of information or some knowledge to try and answer a question, they can say our AI will say something like, hold on a moment or just a second, while sort of completing some of those slower operations and then coming back with, like, a complete response. A couple other things is just I mean, in general, with any kind of automation, being able to plan for the failure case. So, maybe you need an order number to look up the status of an order. Maybe the customer can't find it in their email. In this case, you might have to make sure that you're sort of identifying all of those situations and properly sort of escalating to the right agent if necessary. And then I think one more is just, like, validating that data is captured accurately. So over the phone, there's just, like one one thing we'll do is have the AI agent basically, like, repeat back certain pieces of information to verify that they're sort of captured correctly, before using them in, like, an API call, making sure the request is successful. Fantastic. And our next question here is, how have you been measuring or thinking about quality of the AI portion of the interaction? Good question for us. Oh, go ahead, Leslie. Yeah. Yeah. I can talk about this. So, typically, like, we we'll know, if they're they're calling back again. Right? Because if the customer is not getting their solution that they want or they're not hearing what they want and this this could happen with human agents too. They will call back again. So it's really we're looking at that containment rate and also, like, the sentiment analysis, to make sure that they're not they're they're they're not leaving the call, like, more frustrated than when they started. We were QA ing QA ing it the same way that we would with a human agent. I really don't like, we're like, we don't have to do anything differently. Like, if, you know, like, if the human agent is accurate, if the the AI agent is accurate, if they responded well to the customer, it's we're kind of applying the same metrics there. Yeah. I would probably second that as well. Across channels, how we think about quality is the same kind of rubric any QA team would go through, whether it's an email or a chat or a phone call. Did the agent say the right things? All those kind of, scores. We have a sort of relevant follow on question here. Does Assembled have an AI quality assurance system? So, yeah, we have a number of tools that are meant to be able to sort of improve how you can understand and monitor quality. So we will give sort of interaction level visibility into every AI agent, interaction. So you'll see a full transcript of the conversation and also be able to score those transcripts using the rubric that May and Leslie were talking about directly in the tool. We also have reporting built off of this, and also the ability to just test everything really deeply integrated into the platform. So if you have a specific workflow or, sort of knowledge responses, routing rules you wanna set up, it's built into the platform to be able to test all of that stuff. And so the goal is to give you sort of the testing tools and then both the interaction level and high level recording you need to monitor and improve quality over time. It's usually a bit of an iterative process, and sort of capturing examples where the AI might have done something wrong. We can use those to just generally, improve quality and gather those examples and and get the right behavior in every sort of case. Yeah. Great. So our next question here, for voice AI that results in a handover, does the AI system inform the human agent of the interaction before completing the transfer to avoid customers needing to repeat themselves? Does the AI agent continue to monitor the call for data capture, transcription, etcetera? And I believe that is a a yes, but I will hand it to Sam for the details. Yeah. So short short answer is definitely yes. When the call is transferred to an agent, we basically create a ticket with, a ton of different handoff data. This can be customized sort of down to the issue type level. But, generally, this summary will arrive immediately when the agent picks up the phone with the rest of the ticket, and it'll have, like, an AI generated summary of the caller's issue, the specific type of issue they're calling about, and we can do things like potential next steps or structured data, that's just, like, relevant to follow-up following up on the case. So things like name and email usually, but also, any kind of, like, case specific information as well. Great. And our next question here, do you feel is more useful for GenAI and voice, knowledge based optimization for AI or directly creating coaching and scenarios for specific use cases? I have, I can jump in on this one as well. Mhmm. But it my short answer is that it depends on your business, but most of the time, it's both. So we do have some customers who have, I don't know, a really large percentage of, support cases are related to, like, a single type of issue. And in that case, you can get a lot of leverage from just, like, one really well defined workflow. But usually, the combination of that plus knowledge answers, is the right way to get to sort of meaningful resolution rates. Great. And another question here is more specific to Assembled. Does Assembled have a way of gathering qualitative data and sentiment? Yes. We'd love to, yeah, expand on that for us as well, Sam. Thanks for taking it. By qualitative data, I think, like, there are a lot of just automatically sort of detected attributes of the call. For example, like, what category of issue is this? But, also, if you share more detail, I'm happy to get more specific. Sentiment detection is another thing that AI is also really good at out of the box, so something that we can do as well. Great. Another one here. Happy to see so much engagement. Thanks all. What full resolution rate is desired for a go live on voice AI to be worth the potential friction point for customers who will be routed through topics that cannot be automated? We understand that question here. I'm happy to answer that one. I think the actual the goal kind of of AI handling your cases and that voice AR is they don't have to go through that traditional IVR that might talk about topics that aren't relevant. So the very first question is often kind of what can I help you with and immediately getting into that customer's concern? Is that something we want to immediately pass to an agent? We might collect some of that agent, information, but then pass to an agent versus if it is a topic that can be automated. In those cases, then the agent the AI agent, will actually help the customer in those cases. So, really, what the voice AI agent allows you to do is, remove that sense of friction since you can route directly off of that customer's first sentence on why they're calling and what their pain point is. Great. And a question for Leslie here. Leslie, you mentioned using Maestro QA. Do you also do QA for the AI systems or do you only use the QA scorecards for human agents? For the QA scorecard, we're we're we're using AI in two different ways. For the QA scorecard, the intention is to coach the the agent. And for us, like, we're looking at, are they matching the customer's energy, all of that stuff versus with the QA for for our assembled agent, we're just using the QA quality page for. So the same scorecard doesn't apply to both, but we don't think that the AI agent would have the same issues that, like, a human agent would. 100%. Alright. An assembled question here. How do you manage system outages, or does your system have some sort of monitoring tool to monitor the AI performance? I can take this one. So, there's a couple of different things here. One, we really invest a lot in uptime. So, the I actually don't think we've had a production, outage yet for voice, but keeping the, uptime really high is something we're very focused on. We do have automated monitoring and learning, so we will know instantly if anything, drops. And then another thing we've we've done is also just having, like, the right fallback measures. So if a call gets transferred to us and we're down, we'll make sure that it immediately gets sort of transferred back into your IVR system, and we can even sort of do, like, a health check, to just make sure we're up before transferring out to us and then, like, proactively sort of not transfer calls out to the AI agent and just send them straight into the sort of traditional IVR or support queues. Great. And final question here for Leslie. This is a good one. What is one thing you wish you would have known before implementing AI for telephony? Before implementing AI, I I'm not sure. I I guess, like, the answer to this is, like, I wish that I knew it was going to be this easy. I probably would have done it sooner or that it could apply to so many use cases and fix, like, all of these problems that we're having, especially when I mentioned earlier the contact driver problem. We we were trying to solve that for almost, like, two years when there was, you know, like, a very easy solution that we found with AI. Love that answer. It was a great one. Fantastic. Well, all our questions we have today. So to wrap up here, thank you so much, Leslie, for joining us. Thank you everyone for attending, for all your engagement, for all your very thoughtful questions. We will follow-up with a recording of this webinar. If you'd like to learn more about Assembled approach to voice AI or get a demo where you could actually hear our voice AI agent live, the dynamic lifelike voice agent, and learn more about how it could apply to your use case. You can get a demo by either following up with, the team that will reach out to you after this webinar. You can always check our website, or you can follow the link here in the chat. But either way, we will be following up with you, and we really appreciate your time today. Thank you again to Leslie and to everyone who attended. Hope you have a great rest of your week all. Thanks.